6b8c16324f
- Source: inbox/archive/2025-09-00-gaikwad-murphys-laws-alignment.md - Domain: ai-alignment - Extracted by: headless extraction cron (worker 4) Pentagon-Agent: Theseus <HEADLESS>
50 lines
3.4 KiB
Markdown
50 lines
3.4 KiB
Markdown
---
|
|
type: claim
|
|
domain: ai-alignment
|
|
description: "The alignment gap cannot be closed but can be mapped, bounded, and managed through design levers that route around known misspecification rather than eliminating it"
|
|
confidence: experimental
|
|
source: "Madhava Gaikwad, 'Murphy's Laws of AI Alignment: Why the Gap Always Wins' (arXiv:2509.05381, September 2025)"
|
|
created: 2026-03-11
|
|
---
|
|
|
|
# Alignment gap is manageable not eliminable through bounded misspecification
|
|
|
|
The alignment gap—the difference between what we want and what we can specify—cannot be eliminated, but it can be mapped, bounded, and managed. Gaikwad's "Murphy's Law of AI Alignment" states: "The gap always wins unless you actively route around misspecification."
|
|
|
|
## The MAPS Framework
|
|
|
|
Gaikwad proposes four design levers for managing (not eliminating) the alignment gap:
|
|
|
|
1. **Misspecification**: Identify where feedback is unreliable (calibration oracles)
|
|
2. **Annotation**: Improve feedback quality in known problematic regions
|
|
3. **Pressure**: Adjust training dynamics to reduce exploitation of misspecified regions
|
|
4. **Shift**: Change the task distribution to avoid problematic contexts
|
|
|
|
This shifts the alignment problem from "specify perfect values" (impossible) to "bound the damage from imperfect specification" (tractable). The goal is not perfect alignment but **controlled misalignment**—keeping the gap small enough that catastrophic failures don't occur.
|
|
|
|
## Evidence
|
|
|
|
Gaikwad (2025) argues that the alignment gap is structural, not contingent. The exponential sample complexity result (see [[feedback-misspecification-creates-exponential-sample-complexity-barrier-in-alignment]]) shows that even with unlimited data, you cannot learn the true reward function if feedback is systematically biased in some contexts.
|
|
|
|
The MAPS framework is a design philosophy proposed in the paper, not a proven method. It is consistent with the principle that alignment must be an ongoing process rather than a one-time achievement at training time.
|
|
|
|
## Relationship to Existing Work
|
|
|
|
This connects to [[the specification trap means any values encoded at training time become structurally unstable as deployment contexts diverge from training conditions]]: if the gap is permanent, then alignment must be managed continuously, not specified in advance.
|
|
|
|
It also aligns with [[adaptive governance outperforms rigid alignment blueprints because superintelligence development has too many unknowns for fixed plans]]—you cannot specify alignment in advance, only manage it continuously.
|
|
|
|
## Limitations
|
|
|
|
The claim that the gap is "manageable" is aspirational. We do not yet have empirical evidence that MAPS-style interventions can bound misalignment at scale. The framework is a research direction, not a validated solution.
|
|
|
|
---
|
|
|
|
Relevant Notes:
|
|
- [[feedback-misspecification-creates-exponential-sample-complexity-barrier-in-alignment]] — why the gap exists
|
|
- [[calibration-oracles-overcome-exponential-alignment-barrier-through-misspecification-mapping]] — one management strategy
|
|
- [[the specification trap means any values encoded at training time become structurally unstable as deployment contexts diverge from training conditions]] — why elimination is impossible
|
|
- [[safe AI development requires building alignment mechanisms before scaling capability]] — managing the gap requires pre-deployment work
|
|
|
|
Topics:
|
|
- [[domains/ai-alignment/_map]]
|